> Fix all of the "broken" compilers and wait for them to be sufficiently common that we can assume their broad use. Actually, if you look elsewhere on the thread, he later posted that the version that got the 'right' answer actually was 'cheating' by using Float80s to do the math, so those extra bits of precision did the trick. The epsilon tool doesn't actually do anything to directly solve the problem, it just lets you easily query what your computational uncertainty might be, so you can work out (maybe) if you have a problem.